Proactive Cooling Control Algorithm for Data Centers Based on LSTM-Driven Predictive Thermal Analysis

Overview

The study starts from the problem that reactive cooling strategies built on static thresholds are inadequate for dynamically changing heat loads, causing energy inefficiency and local hot-spot risk. The test room is 15 × 10 × 3.5 m with 20 racks in 2 rows, 100 kW average and 150 kW peak IT load, two chilled-water CRACs with N+1 redundancy and hot/cold aisle containment. There are 150 sensors (6 per rack at top, middle and bottom of inlet and outlet; 4 at the CRACs; 26 elsewhere), with data from 1 January to 30 April 2025.

The LSTM uses a 60 min input window and a 30 min prediction horizon, with a reported RMSE of 0.25 and R² of 0.985. Prediction stayed stable down to two sensors per rack (one inlet, one outlet); below that, RMSE rose by about 22%. Median R² across 120 sensors exceeded 0.97 over the four months. The forecast drives pre-emptive changes to CRAC fan speed and valve position instead of reactive PID control.

In simulation against PID control on historical data, mean PUE fell from 1.48 to 1.30, which the authors express as a 37.5% cut in cooling energy and a 12.16% cut in PUE. Hours above 27 °C fell from 48 to 2.

Limits: the energy comparison is a data-driven simulation on four months of historical data, not a live A/B test; the cooling response was simulated with a thermodynamic model driven by real operating data, not measured in a controlled field trial. The most transferable result is the sensor-density sensitivity check: one inlet and one outlet sensor per rack is the minimum for reliable prediction.

Key points

1. Test room: 15 × 10 × 3.5 m, 20 racks in 2 rows, 100 kW average and 150 kW peak IT load, two chilled-water CRACs with N+1 redundancy and hot/cold aisle containment.
2. 150 sensors (6 per rack at top, middle and bottom of inlet and outlet; 4 at the CRACs; 26 elsewhere), with data from 1 January to 30 April 2025.
3. The LSTM uses a 60 min input window and a 30 min prediction horizon, with a reported RMSE of 0.25 and R² of 0.985.
4. Prediction stayed stable down to two sensors per rack (one inlet, one outlet); below that, RMSE rose by about 22%; median R² across 120 sensors exceeded 0.97 over four months.
5. In simulation against PID control on historical data, mean PUE fell from 1.48 to 1.30 (a 37.5% cut in cooling energy and 12.16% cut in PUE); hours above 27 °C fell from 48 to 2.
6. The cooling response was simulated with a thermodynamic model driven by real operating data, not measured in a controlled field trial.

Using it in design

• For predictive cooling control, plan at least one inlet and one outlet temperature sensor per rack; below that, prediction error rose by about 22%.
• Do not commit to the PUE gain of proactive control (1.48 to 1.30) without live trial validation; the result rests on a four-month historical-data simulation.
• Define CRAC fan-speed and valve changes as pre-control (30-minute prediction horizon) and test it also under N+1 maintenance conditions.

Who should read it

Engineers developing data-centre cooling control, sensor placement and predictive control algorithms.

Document contents

1. 1. Introduction
2. 2. Methodology (2.1 System Architecture, 2.2 Data Acquisition and Preprocessing, 2.3 Predictive Modeling and Digital Twin Construction, 2.4 Control Logic and Security Constraints, 2.5 Comparison with MPC and RTDO)
3. 3. Experiment and Implementation (3.1 Operating Conditions: baseline load, humidity stability and control, peak load, dynamic load fluctuation, equipment maintenance condition (N+1 redundancy test); 3.2 Control Logic and Execution; 3.3 Implementation and Reproducibility Details)
4. 4. Results and Discussion (4.1 Temperature Prediction, 4.2 Hotspot Suppression Effect, 4.3 Energy Consumption, 4.4 Dynamic Response of CRAC Operation, 4.5 System Robustness, 4.6 Ablation and Sensitivity Analysis, 4.7 Economic and Practical Considerations, 4.8 Robustness and Sensitivity Check)
5. 5. Conclusions

Related standards

• ASHRAE TC 9.9 Thermal Guidelines for Data Processing Environments
• ISO/IEC 30134-2 (PUE)

Read the document

Publication details

Authors

Jieying Liu; Rui Fan; Zonglin Li; Napat Harnpornchai; Jianlei Qian

Publisher

Applied System Innovation (MDPI AG)

Year

2026

Volume

9

Issue

1

Pages

21

Document language

English

Source type

Scientific

Access

Open licence — downloadable

DOI

10.3390/asi9010021

Licence

CC BY 4.0 — share with attribution

Licence evidence

www.mdpi.com

Citation